Integration of multiscale simulations and machine learning for predicting dendritic microstructures in solidification of alloys

Abstract

This study presents an integration of machine learning (ML) with a multiscale computational framework to predict primary dendrite arm spacing (PDAS) during alloy solidification. Analytical models, such as Hunt (HT) and Kurz-Fisher (KF), provide the basis for developing parametric and non-parametric ML models that capture the influence of processing conditions and material properties on PDAS. The training and testing dataset is generated from high-throughput phase-field simulations across various alloy systems, incorporating material properties calculated via molecular dynamics. While non-parametric models, such as decision trees, random forests, and gradient boosting decision trees, perform well in training, they encounter overfitting challenges due to the limited size of the computational dataset. In contrast, parametric models, including linear, ridge, and lasso regression, successfully capture key PDAS features, producing predictions that align closely with experimental data. Overall, parametric ML-based models show a stronger dependence on pulling velocity, temperature gradient, and material properties compared to the HT and KF models, offering a more accurate tool for predicting PDAS and optimizing alloy solidification processes.